UC Irvine

Low Earth orbit (LEO) satellites inherently possess desirable attributes for navigation: (i) abundance, (ii) geometric and spectral diversity, and (iii) high received powers. However, the first prerequisite to satellite navigation is to know the satellites’ ephemeris (i.e., position and velocity over time) and clock error states. Unlike global navigation satellite systems (GNSS), specifically designed for navigation, with satellites in medium Earth orbit (MEO) that constantly transmit ephemeris and clock corrections to users in their signals, LEO satellites, mainly operated by private companies, generally do not openly send such information in their proprietary signals. The quality of oscillators on-board LEO satellites’ as well as their clock error states are completely unknown. Moreover, the most accurate publicly available information on LEO satellites’ ephemerides is in the form of two-line element (TLE) files, which yield ephemerides with errors of a few kilometers in position and a few meters per second in velocity. Consequently, LEO satellites’ states are completely unknown (clock errors) or uncertain at best (ephemeris).

This thesis addresses the aforementioned challenges by performing the opportunistic estimation of LEO satellites’ states. First, a study of the use of machine learning for satellite orbital determination is conducted. Multiple models for orbit propagation are analyzed, and the model with the best performance is found. The model is then utilized in a STAN framework, experimentally demonstrating its capability of producing satellite ephemeris good enough to allow for desirable navigation solutions. Next, a framework is proposed to collect training data when the target ephemeris data is not available due to satellites not transmitting their ephemeris. Finally, the framework’s feasibility is demonstrated experimentally first by localizing a stationary receiver and second by coupling an IMU with LEO observables to navigate a moving ground vehicle.